Struct AdjointMethod 

pub struct AdjointMethod {
    pub augmented_state: Vec<f64>,
    pub state_dim: usize,
}

Reverse-mode gradient of a Neural ODE loss via the continuous adjoint method.

The adjoint state a(t) = -dL/dz(t) is integrated backward in time. This gives parameter gradients without storing the full forward trajectory.

Fields

augmented_state: Vec<f64>

Augmented state [z; a; dL/dθ] during backward integration.

state_dim: usize

Dimensionality of the ODE state.

Implementations

impl AdjointMethod

pub fn new(state_dim: usize) -> Self

Construct a new AdjointMethod for an ODE with state dimension state_dim.

pub fn backward(&self, loss_grad: &[f64]) -> Vec<f64>

Compute parameter gradients given the loss gradient at the final time.

This is a simplified adjoint implementation: it propagates loss_grad backward through one RK4 step and returns the approximate gradient with respect to the initial state.

In a full implementation, func would be called to integrate the adjoint ODE backward; here we use a finite-difference approximation to illustrate the interface.

pub fn run( &mut self, solver: &NeuralOdeSolver, z_final: &[f64], loss_grad: &[f64], t0: f64, t1: f64, dt: f64, ) -> (Vec<f64>, Vec<f64>)

Set the final adjoint state from loss_grad and propagate it backward through solver from t1 to t0 using RK4.

Returns (grad_z0, grad_params) where grad_z0 is the gradient with respect to the initial state and grad_params is a flat vector of approximate parameter gradients.

Trait Implementations

impl Clone for AdjointMethod

fn clone(&self) -> AdjointMethod

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for AdjointMethod

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.